pkgsrc wrapper script

rvp

pfr This script empowers you [...]

"Empowers you ..."?! Empowers -- Aw, @Jay-sus H!

#!/bin/sh
[...]
selected_pkg=$(fzf --preview 'cat $PKGSRCDIR/{}/DESCR' <<< "$pkgs")

So, A.I., all /bin/shs understand here-strings, eh? Empowers, my furry POSIX ...!

pfr

rvp "Empowers you ..."?!

The bots do like to exaggerate! (these aren't my words)

rvp

#!/bin/sh
[...]
selected_pkg=$(fzf --preview 'cat $PKGSRCDIR/{}/DESCR' <<< "$pkgs")

So, A.I., all /bin/shs understand here-strings, eh?

Can you elaborate on what is wrong here? Will this not work under /bin/sh? Is there a more POSIX way of writing this part?

pin Note, I have nothing against you trying to automate the process but, I would recommend learning a bit further before doing so. It will save you future headaches when the automation fails and you have no clue why.

I agree, hence why I posted it here first instead of blindly running this script. I acknowledge the potential for issues when doing these types of things, and as @rvp has already pointed out, the LLM's are great at seeming like they're always correct but often aren't.

pfr distclean: Removes everything related to the build process, including downloaded archives.

Wrong, it cleans only downloaded stuff for the package currently being built. It doesn't clean downloaded files for any of the built dependencies.

Also, stupid to do if you're building things that are separated packages from the same source. For example, LLVM contains libunwind, libcxx and libcxxabi, as well as local copy of cmake. This means that, if you run distclean after building LLVM and then you want to build libunwind, it will simply have to fetch the sources again, as you just removed them.

Note, I have nothing against you trying to automate the process but, I would recommend learning a bit further before doing so. It will save you future headaches when the automation fails and you have no clue why.

Or, maybe ignorance is bliss after all.

rvp

pfr Is there a more POSIX way of writing this part?

Let's see if the A.I. can do it first. Tell it that a) <<< is not POSIX, and b) you'd like it to handle the case of fzf returning multiple selections for steps 7. & 8.

pfr

rvp

This any better?

#!/bin/sh

# Check if fzf is installed
if ! command -v fzf >/dev/null 2>&1; then
  echo "Error: fzf is required. Please install it first." >&2
  exit 1
fi

# Check if fd-find is installed
if ! command -v fd >/dev/null 2>&1; then
  echo "Error: fd-find (fd) is required. Please install it first." >&2
  exit 1
fi

# Check if bat is installed
if ! command -v bat >/dev/null 2>&1; then
  echo "Error: bat is required. Please install it first." >&2
  exit 1
fi

# Path to the user's pkgsrc directory
PKGSRCDIR="$HOME/pkgsrc"

# Check if the pkgsrc directory exists
if [ ! -d "$PKGSRCDIR" ]; then
  echo "Error: $PKGSRCDIR directory not found. Please create it and place your pkgsrc directory inside." >&2
  exit 1
fi

# Function to show package description
show_description() {
  pkg="$1"
  descr_file="$PKGSRCDIR/$(dirname {})/$pkg/DESCR"
  if [ -f "$descr_file" ]; then
    bat "$descr_file"
  else
    echo "No description available for $pkg"
  fi
}

# Find available packages recursively in all category directories and display selection menu
pkgs=$(fd --type d --min-depth 4 --max-depth 4 'Makefile' "$PKGSRCDIR" | xargs -I{} basename {})
selected_pkg=$(printf '%s\n' "$pkgs" | fzf --preview 'show_description {}' --preview-window=right:70% --prompt="Select a package: ")

# Check if a package was selected
if [ -n "$selected_pkg" ]; then
  # Display package description and install
  show_description "$selected_pkg"
  cd "$PKGSRCDIR/$(dirname {})/$selected_pkg" && make clean install
else
  echo "No package selected"
  exit 1
fi

rvp

pfr This any better?

Makes my eyes glaze. I think A.I. should stick to writing milquetoast press-releases. Doesn't mean we can't teach it any better:

(warning: untested, since I don't have fzf nor bat installed)

#!/bin/sh

set -eu

die() {
	echo >&2 "$0: $*"
	exit 1
}
getkey() {
	local orig=$(stty -f /dev/tty -g)
	stty -f /dev/tty -icanon min 0 time 20		# 2 secs.
	key=$(dd if=/dev/tty bs=1 count=1 2>/dev/null)
	stty -f /dev/tty "$orig"
	echo "$key"
}

for C in fzf bat
do	command -v "$C" >/dev/null || die "$C: reqd. cmd. not found"
done

D="$HOME/pkgsrc"
cd "$D"		# || exit 1

# Only category/package considered
#
P=$(printf '%s\n' */*/Makefile | sed 's;/Makefile$;;')
[ "$P" = "*/*" ] && die "$D: not a pkgsrc base dir."

for P in $(echo "$P" | fzf --preview 'bat {}/DESCR')
do	echo "====> Going to build & install: $P (q to QUIT)"
	[ "$(getkey)" = q ] && break
	make -C "$P" install clean
done

Learn, little A.I., learn!

nettester

Is that again made by ChatGPT? I wonder what's up with all those {}. What's the point in something like basename {} and the $(dirname {}) part (which i think just statically expands to .) of the directory being switched to seemingly makes even less sense (not to mention that make has a -C switch, which could fully avoid the need to cd).

No offense but in my opinion trying to get those language models to program stuff is not that great of an idea. They don't really know what they are doing and even if they might get it right sometimes picking out the subtle errors of some source one isn't familiar with is almost worse than just writing the whole thing from scratch.

Edit: I also don't understand why it insists on using fd instead of boring old find which seemingly would do pretty much the same thing and likely could also avoid xargs entirely.

remiliascarlet

nettester Is that again made by ChatGPT? I wonder what's up with all those {}.

Imposter syndrome is very real, isn't it?
Pre-ChatGPT, new coders climbed Mount Stupid, fell off that mountain, and found their path to enlightment.
Post-ChatJeopardy, new coders climb Mount Stupid, and get stuck on there forever.

rvp Learn, little A.I., learn!

Let's call it for what it really is.
It's not an AI, but a bullshit generator.

pfr

rvp I like this very much! 🙏

Tested, works perfectly. Damn, AI has a ways to go huh.

kc9udx

remiliascarlet I always call it Fake Intelligence but you might have it better.

I think it should be called Deception Mastered, or Mastered Deception. But that just doesn't sound as nice.

rvp

remiliascarlet It's not an AI, but a bullshit generator.

Recalls to mind Frankfurt's distinction between liars and bullshitters: Liars have to care about the truth (in order to lie better), whereas bullshitters just don't care.

A.I.s, at present, are distinctly nearer to the smelly side of that line.

I think that part of the reason for this tendency to bullshit is simply due to the way A.I.s work right now: They start at the base (the initial try at a "goal") then always work up the "tree" of possibilities when you say "Fail! Try again!", instead of returning to the base of the tree to make a fresh cast--like a human would after a period of failures. Maybe it just doesn't want to throw away all the "data" it has accumulated along its search path. That would account for what we saw just now with this script: the more you push it, the more unmoored from reality it got (compare the initial offering with the final script).

nettester

[unknown] You are onto something there i think 😉

https://invidious.lunar.icu/watch?v=FojyYKU58cw

I 100% guarantee that if you know just a little bit of chess and haven't seen this yet you'll be absolutely mindblown 😃

kc9udx

nettester I laughed way too hard at that. But who calls a knight a horse? 🤣🤣

BTW what happened to my previous reply? 😳

nettester

kc9udx I'm not sure what's customary in this regard in English (probably quite obviously not a native speaker) but at least in German calling the Knight Pferd (horse) is not that uncommon. Actually the term Ritter (which would be the direct translation of Knight) is not used at all and it's rather called Springer (jumper). I don't know the guy but i guess if he isn't a native speaker and would usually speak a language where using horse is common getting carried away and forgetting to adjust the term to English probably could happen quite easily.

kc9udx

Historically, the word knight is actually the word Knecht 🙂 (German to Old English and then corrupted over time). I sometimes call it that. But then I often call jacks Knaves, too.

BTW I often hear novices calling them "horsey". But didn't expect "horse" from someone really serious.

I wonder who can delete my erroneous extraneous replies above.

kc9udx I wonder who can delete my erroneous extraneous replies above.

Fixed 😉

pfr

rvp I presented your working script back to it and rather than admit defeat, it once again attempted to improve on it by simply regurgitating nearly the exact same script or had given me previously, which was still miles away from actually working.

I told it, this script right here, it works, learn from it. And it assured me it would 🤣

rvp

pfr Maybe using chatbot A.I.s to write shell-scripts is an inherently silly idea (other than for the lulz)... just like using it to play chess (as a chess IM, surely GothamChess has heard of AlphaZero).

nettester

rvp The problem is that a lot of people are using it that way while being 100% oblivious to the fact that it's inherently silly. The defense i usually hear is something along the lines of "I'm just using it to get a general idea of how a solution would look like." which doesn't hold water in my opinion as the time spent on figuring out if or how far whatever the AI comes up with is actually valid would usually also suffice to just read the relevant documentation or research actual examples that are proven to be valid or at least have a high likelihood of that.

To make matters worse people going at it this way often times also flat out lack the capability to fact check the result. I predict that technicians of the not to far future will have to deal with quite bit of bizarre errors introduced by AI written code/configuration that managed to superficially work but was still broken overall.

rvp

nettester The problem is that a lot of people are using it that way while being 100% oblivious to the fact that it's inherently silly.

I should've qualified:

Specialized A.I. works (eg. AlphaZero).
LLMs also work--provided you
a) train them hard (eg. Grammarly, Github Co-Pilot), or
b) use it for is non-mission-critical stuff: for eg. generating pix of fantasy ladies (don't look too closely at their hands), or generating music.

Co-Pilot--which is not free--is trained on billions of lines of Python and Go code. How many lines of shell-script has ChatGPT been fed? Doubt it's anywhere close to a billion. More likely just a lot of stuff randomly scraped from the Internets. And I bet most of those were bash or zsh scripts--which is why A.I. thought /bin/sh should understand <<<.

nettester

rvp I should've qualified

Well, i think you pretty much did that by specifying chatbot AI. Sure, you can probably train specialized AI's to write code. No doubt about that but i think it's a long, long way before it'll actually be trustworthy to write longer portions without close supervision and it will always (well, at least for the foreseeable future) somewhat choke on complex problems regarding more exotic topics as it can only rely on existing data but not do any research of it's own let alone actually understand concepts. Besides actually defining complex problems well enough to get some clear target for an AI to work towards for producing an useful solution seems to be quite a hurdle.

I guess it will become very good at writing common code snippets though. I'm seriously looking forward to just hand some AI a struct definition and have it write the boring command line parsing logic that fills it 😉

Actually the chess video, as silly as it is, can be seen as an example of the limitations of AI. I mean, one just has to look at how both engines start off quite reasonable and get more and more eccentric the further the game progresses. The reason should be pretty obvious: The more complex the game state gets the less data they have to draw from pushing them to make decisions based on states that are further and further away from actual reality reducing the chances for them to make sense.

Sure, a specialized AI might be able to interpolate and deduct from multiple somewhat close data points but they still don't understand concepts or are able to actually invent things as they are basically parrots on steroids and as soon as they are confronted with something they can't simply copy it's down to their interpolation abilities.

Edit: Yeah, probably kind of like trying to interpolate a hand without understanding the concept of a hand.